Interpretable Machine Learning for One-Year Survival Prediction in Hepatocellular Carcinoma: A SHAP- and LIME-Based Explainable Artificial Intelligence Framework on the UCI HCC-Survival Cohort

Objectives: To develop and interpret machine learning (ML) models for predicting one-year survival in patients with hepatocellular carcinoma (HCC), and to characterize model behavior using an explainable artificial intelligence (XAI) framework combining global and local interpretability with calibration and clinical-utility analysis.
Methods: This was a secondary analysis of the publicly available, de-identified University of California, Irvine (UCI) HCC-Survival da-taset (165 patients; 49 features; outcome, one-year survival coded as 1 = alive and 0 = deceased). Missing data (10.2% overall) were handled with multivariate imputation by chained equations (MICE), and class imbalance (102 alive vs. 63 deceased) was addressed with the synthetic minority over-sampling technique combined with Tomek links (SMOTE-Tomek); all preprocessing was performed within cross-validation folds to prevent leakage. Seven ML algorithms-logistic regression, random forest, extreme gradient boost-ing (XGBoost), light gradient boosting machine (LightGBM), categorical boosting (CatBoost), support vector machine (SVM), and k-nearest neighbors (KNN)-were trained under 10-fold stratified cross-validation. SHapley Additive exPlanations (SHAP) provided global interpretability and Local Interpretable Model-Agnostic Explanations (LIME) provided local interpretability. Platt scaling was used for calibration and decision curve analysis (DCA) for clinical utility.
Results: Random forest achieved the best discrimination (area under the receiver operating characteristic curve [AUC] = 0.808 ± 0.083; accuracy = 0.746; F1 = 0.716; Brier score = 0.176; Matthews correlation coefficient [MCC] = 0.446), followed by XGBoost (AUC = 0.795) and LightGBM (AUC = 0.783). SHAP analysis identified alpha-fetoprotein (AFP), alkaline phosphatase (ALP), serum iron, performance status, and hemoglobin as the most influential predictors of one-year survival, consistent with prior findings on this cohort. LIME provided con-cordant patient-level explanations. DCA indicated a positive net clinical benefit across a clinically relevant range of risk thresholds.
Conclusion: On a small, single-center cohort, an explainable ML framework yielded moderate but robust discrimination and, more importantly, transparent and clinically coherent explanations. Rather than maximizing performance alone, the framework empha-sizes interpretability, calibration, and clinical utility, supporting its potential role as a transparent decision-support aid that warrants validation in larger, prospective cohorts.

Keywords: Hepatocellular carcinoma, machine learning, artificial intelligence, risk prediction, clinical decision support systems, prognosis